Living with Generalized Linear Mixed Models

In the 1980s, before PROC MIXED or PROC GENMOD, "linear models" meant the "general" linear model as implemented by PROC GLM. Three decades later, the meaning of "linear models" has fundamentally changed. The introduction of PROC GLIMMIX in 2005 was a watershed moment. Now "linear model" means "generalized linear mixed model." The notion that the "general" linear model was once considered "general" seems quaint. In living with PROC GLIMMIX over the past five-plus years, several issues have become apparent that were not issues during PROC GLM's heyday. One set of issues concerns inference space and conditional versus marginal modeling. When you learn about the GLMMs, you often find that you must first "unlearn" the old ways of thinking about models. The second set of issues concerns power, sample size, and planning. These areas have not caught up with GLMMs. Finally, applied statistics curriculum seems caught in a time warp, leaving students and especially consumers of statistical methods unprepared for contemporary statistical practice.